feat(inbox): match non-invoice documents via prominent amounts (#2048)

* feat(inbox): match non-invoice documents via prominent amounts

Bankintyg, bank agreements and other documentKind "other" PDFs carry no
invoice-style total, so extraction correctly left totals.total null and the
document became structurally unmatchable: findUnderlagCandidates hard-drops
items without a comparable amount and the picker lost the 40% amount signal.

- extraction: new prominentAmounts[] field (amount + document's own label),
  populated only when totals.total is null; account/org/phone/reference
  numbers and zero amounts excluded. totals.total semantics untouched.
- matching: bestProminentAmountVariance() tries each printed amount and
  feeds calculateMatchConfidence at reduced weight (0.3 vs 0.4) in both the
  agent candidate scorer and TransactionMatchPicker.
- UI: inbox rail shows the detected amounts read-only for such documents,
  list falls back to a single distinct prominent amount, and extraction no
  longer reads as "found nothing".

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Hqm9QgdyNAFaWiz6Ww7pgb

* fix(inbox): discount prominent-amount fallback instead of reweighting it

Skeptic pass refutations on the first commit: normalized weighting made a
reduced amount weight self-defeating. Date + exact fallback amount with no
merchant scored (0.25+0.3)/0.55 = 1.0 ("100% sakerhet" on a wrong same-day
transaction), and a DISAGREEING fallback amount scored above a disagreeing
invoice total (0.67 vs 0.60) because shrinking the weight also shrank the
penalty.

- score fallbacks at full amount weight, then multiply by a flat
  FALLBACK_CONFIDENCE_FACTOR (0.85): agreement caps below certainty,
  disagreement stays at least as damning as for a real total.
- agent candidate surface additionally requires the document date within
  DATE_TOLERANCE_DAYS, so an avtal listing 349 kr no longer matches every
  future 349 kr charge from the same counterparty.
- bestProminentAmountVariance returns which amount matched + its document
  label, and the match reason names it ("Exakt belopp i dokumentet: 2 500
  SEK (Engangspris)"): no more bare "Exakt belopp" reaching the agent while
  total_amount is null.
- prompt: prominentAmounts restricted to non-invoice documentKinds, and
  never a parking spot for an unreadable invoice total.
- fix the stale "deliberately the same list" comment on
  EXTRACTED_FIELD_ACCESSORS.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Hqm9QgdyNAFaWiz6Ww7pgb

* fix(receipt-hunt): never propose on the prominent-amounts fallback

Second skeptic pass: the nightly hunt is a third consumer of
scoreUnderlagCandidates and inherited the fallback unaware. A bankintyg
whose printed "Insatt belopp" equals a same-day outflow scores 0.85, which
clears CERTAIN_CONFIDENCE (0.8) and skips LLM adjudication, on a pairing
wrong by construction (the hunt scans outflows only; "Insatt belopp"
labels an inflow), with document_amount null in the approval preview.

UnderlagCandidate now carries amountSource ('total' | 'prominent') and
selectProposals drops fallback-scored candidates. Non-invoice documents
stay reachable through the manual picker and the agent candidate surface,
both of which have a human reading the amounts.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Hqm9QgdyNAFaWiz6Ww7pgb

* fix(inbox): round fallback confidence via roundOre, not the naive pattern

The two confidence discounts (and their test) tripped the naive-ore-round
antipattern ratchet (625 vs baseline 622); use the sanctioned helper.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Hqm9QgdyNAFaWiz6Ww7pgb

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

DECISIONS.md

@@ -1364,6 +1364,9 @@ One line per decision: `[YYYY-MM-DD] <decision>: <why>`. Appended by agents and
 [2026-08-30] book_skattekonto_row(s) tier 'medium' + scope 'transactions:write': rule-driven booking with no caller-supplied lines mirrors book_mileage_period (not create_voucher's 'high'); scope follows reconcile_residual (books an outside row). Commit service gates on SKATTEVERKET_ENABLED for HTTP-dispatcher parity, recoverable so the op stays pending.
 [2026-08-30] Reminder text overrides (company_settings.reminder_text_overrides, level_1..3 x subject/body): the defaults are expressed as placeholder patterns (REMINDER_EMAIL_DEFAULT_TEXTS) and BOTH the stock mail and overrides render through the same substitution pipeline (applyPlaceholders + escape per output variant), so the settings-UI prefill is byte-for-byte the mail that goes out and cannot drift; this differs from the invoice_email_texts precedent, whose hand-written pattern forms can drift from the coded defaults. The level-3 default body is now an explicit inkassovarning (8 days, fordran till inkasso, costs per lag (1981:739)) but the level TITLE stays 'Slutlig paminnelse': the title is reused as the level name in settings labels and subject prefix, and renaming it everywhere is wording churn beyond the ask. An overridden subject owns the whole line (no automatic ' (inkl. drojsmalsranta)' suffix; {belopp} already includes surcharges), the stock subject keeps the suffix byte-identically. No pg test for the migration: a declarative CHECK (jsonb_typeof object) identical in shape to invoice_email_texts (20260703091000), which also shipped without one. The v1 REST/MCP update_company_settings surface was NOT extended: it is a curated field set with staged operations and its own placeholder refinement, a separate parity slice. typecheck/antipattern baselines deliberately not ratcheted in this diff: both one-count drops predate the branch (main drift), gates only fail on increase.
 [2026-08-30] PR #2021 round 2 (#546): the relayed Peppol buyer restriction now says the customer's org number must not be a personnummer (prepareParty('buyer') in lib/invoices/peppol-bis-billing.ts refuses it with BUYER_PARTICIPANT_IDENTIFIER_UNSUPPORTED, so an enskild firma CUSTOMER is refused, not only an enskild firma sender), Step 4 of the invoicing-rules workflow points at the Peppol section so a top-down reader never reaches the external-provider fallback first, the mark-sent recovery is scoped to the still-draft invoice in every text (INVOICE_MARK_SENT_REPAIR_REQUIRED leaves the invoice sent with the verifikat posted and a second mark-sent returns 409; the reviewer's proposed repair tool gnubok_link_invoice_to_voucher is the PAYMENT link and requires status sent/overdue/partially_paid, so no tool is named and the repair is left to support), and the verifikat parenthetical says "under faktureringsmetoden" (kontantmetod and defer_invoice_booking companies get none at issue). The guard test now also pins the two v1 route descriptions by reading the route source (apiskill:check only detects generated-vs-source drift, not a truth regression). The atom bump was seeded as a THIRD append-only migration (20260830101500, atom v9) rather than consolidating to one: the Supabase preview branch for the PR (xxnqggttsefleehmarjo) has applied both 20260829000100 and 20260829010000 per its schema_migrations, so deleting either would leave a remote with versions absent from the repo, the orphan class the migration rule forbids; all three seeds are idempotent upserts with the version guard, so prod applying them in sequence ends at v9. The generator's max-plus-one name (20260829010001) was renamed to 20260830101500 for the same reason as round 1 (newer than every file on origin/main and every sibling worktree; skills:check hashes content, the pg replay test globs the seed).
+[2026-08-30] Non-invoice amount fallback uses a new prominentAmounts extraction field at reduced match weight (0.3 vs 0.4), not a relaxed totals.total: "total = what the buyer pays" keeps invoice/receipt booking paths unregressed, and one-of-several printed figures agreeing is weaker evidence than a total agreeing. Candidate floor (0.6) deliberately unchanged, so a dateless bankintyg still only surfaces via the manual picker.
+[2026-08-30] Skeptic pass on PR #2048 replaced the fallback's reduced amount WEIGHT (0.3) with a flat confidence FACTOR (x0.85): normalization made the reduced weight both let date+amount-only fallbacks reach 1.0 and score a DISAGREEING fallback above a disagreeing total (0.67 vs 0.60). Fallback candidates additionally require the document date within DATE_TOLERANCE_DAYS on the agent surface, and the match reason names the matched figure + document label since total_amount stays null.
 [2026-08-30] Pre-migration inbox marker cutoff derived at query time from max posted source_type='import' entry_date (excl. series M) instead of a stored sie_imports coverage column: no migration/backfill needed and undo/replace self-corrects; series M excluded because the importer's omforingsverifikation is dated at fiscal year end.
 [2026-08-30] Skeptic round on PR #2047: cutoff armed by a completed sie_imports row (API clients can post source_type='import', which would false-mark never-migrated companies) and omforingsverifikation excluded by description prefix instead of voucher_series 'M' (imported vouchers keep the source file's series; 6 prod companies use series M for real vouchers). Residual known gap: skip-window covered only by the omforing stays unmarked; DuplicateBookingDialog is the backstop.
 [2026-08-30] Correction to the PR #2047 residual-gap note: the DuplicateBookingDialog backstop only covers a single skipped bank movement within 7 days of the omforing's fiscal-year-end date; mid-year and aggregated-skip variants are unmitigated and accepted on rarity (conjunction of skipped voucher + tail date + synced account + fetched history). Exact closure = persist skipped-voucher max date as coverage_end on sie_imports at import time (skippedDetails already has the dates); filed as follow-up.
+[2026-08-30] Receipt-hunt excludes prominent-amounts fallback candidates (amountSource tag on UnderlagCandidate): the nightly hunt scans outflows only and its 0.8 skip-adjudication threshold was calibrated for invoice totals, so a fallback pair (0.85 on date+printed-figure, no merchant) would auto-classify certain on a wrong-by-construction pairing. Fallback docs stay reachable via picker + agent candidates.